| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-f32-vcvt/neon.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2021 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
#include <assert.h> |
| 11 |
|
| 12 |
#include <arm_neon.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/intrinsics-polyfill.h" |
| 16 |
#include "xnnpack/vcvt.h" |
| 17 |
|
| 18 |
|
| 19 |
void xnn_qu8_f32_vcvt_ukernel__neon_u8( |
| 20 |
size_t batch, |
| 21 |
const uint8_t* input, |
| 22 |
float* output, |
| 23 |
const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 24 |
{ |
| 25 |
assert(batch != 0); |
| 26 |
assert(batch % sizeof(uint8_t) == 0); |
| 27 |
assert(input != NULL); |
| 28 |
assert(output != NULL); |
| 29 |
|
| 30 |
const int16x8_t vminus_zero_point = vdupq_n_s16(-params->scalar.zero_point); |
| 31 |
const float32x4_t vscale = vld1q_dup_f32(¶ms->scalar.scale); |
| 32 |
for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) { |
| 33 |
const uint8x8_t vx = vld1_u8(input); input += 8; |
| 34 |
|
| 35 |
const int16x8_t vhx = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(vminus_zero_point), vx)); |
| 36 |
|
| 37 |
const int32x4_t vwx_lo = vmovl_s16(vget_low_s16(vhx)); |
| 38 |
const int32x4_t vwx_hi = vmovl_s16(vget_high_s16(vhx)); |
| 39 |
|
| 40 |
float32x4_t vy_lo = vcvtq_f32_s32(vwx_lo); |
| 41 |
float32x4_t vy_hi = vcvtq_f32_s32(vwx_hi); |
| 42 |
|
| 43 |
vy_lo = vmulq_f32(vy_lo, vscale); |
| 44 |
vy_hi = vmulq_f32(vy_hi, vscale); |
| 45 |
|
| 46 |
vst1q_f32(output, vy_lo); output += 4; |
| 47 |
vst1q_f32(output, vy_hi); output += 4; |
| 48 |
} |
| 49 |
if XNN_UNLIKELY(batch != 0) { |
| 50 |
assert(batch >= 1 * sizeof(uint8_t)); |
| 51 |
assert(batch <= 7 * sizeof(uint8_t)); |
| 52 |
|
| 53 |
const uint8x8_t vx = vld1_u8(input); |
| 54 |
|
| 55 |
const int16x8_t vhx = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(vminus_zero_point), vx)); |
| 56 |
|
| 57 |
const int32x4_t vwx_lo = vmovl_s16(vget_low_s16(vhx)); |
| 58 |
const int32x4_t vwx_hi = vmovl_s16(vget_high_s16(vhx)); |
| 59 |
|
| 60 |
float32x4_t vy = vcvtq_f32_s32(vwx_lo); |
| 61 |
vy = vmulq_f32(vy, vscale); |
| 62 |
|
| 63 |
if (batch & (4 * sizeof(uint8_t))) { |
| 64 |
vst1q_f32(output, vy); output += 4; |
| 65 |
vy = vcvtq_f32_s32(vwx_hi); |
| 66 |
vy = vmulq_f32(vy, vscale); |
| 67 |
} |
| 68 |
float32x2_t vy_lo = vget_low_f32(vy); |
| 69 |
if (batch & (2 * sizeof(uint8_t))) { |
| 70 |
vst1_f32(output, vy_lo); output += 2; |
| 71 |
vy_lo = vget_high_f32(vy); |
| 72 |
} |
| 73 |
if (batch & (1 * sizeof(uint8_t))) { |
| 74 |
vst1_lane_f32(output, vy_lo, 0); |
| 75 |
} |
| 76 |
} |
| 77 |
} |